肺部電腦斷層掃描腫瘤偵測之演算法: 基於卷積神經網路之腫瘤/非腫瘤分類器

Abstract

腫瘤是一群異常細胞群因脫離正常生長調控，進而可以不受限制的快速生長並停留在細胞週期的任意過程，常因細胞凋亡以及端粒酶的機制喪失而不死亡，成為中高收入國家以及高收入國家的主要死亡原因之一。 近年來癌症始終是衛生福利部公布的死亡原因第一名，以105年台灣衛生福利部資料為例台灣惡性腫瘤死亡人數為 47,760 人，占總死亡人數的 27.7％，而2014年二月出版世界癌症報告指出2035年時全球癌症死亡率會增加到2014年的三倍。細分全球惡性腫瘤死亡比率，氣管、支氣管和肺癌類別一向都是前三名，然而，肺癌第三期、第四期的五年存活率都相當低，因此，早期診斷肺癌是延長肺癌存活率的最佳方式。 隨著科技的發展，目前醫學影像器材的選擇也十分多元，如胸部X光、低劑量電腦斷層掃描、正子造影等。胸部X光檢測應用出現時間最早，設備成本以及耗材成本低廉，因此應用也最廣，然而胸部X光檢查無法檢測到直徑小於一公分的腫瘤。即使腫瘤直徑大於一公分，也容易被氣管、支氣管、胸腔、肋骨等遮住而出現檢測盲點，當胸部X光檢測出肺癌時常已經為時已晚。 電腦斷層掃描影像是由X光照射人體，不同組織對於X光吸收率不同，感測器收到的訊號也不同，之後由電腦重建而成的三維灰階影像。相對胸部X光影像，電腦斷層掃描三維影像擁有高解析度，直徑只有1mm到2mm的腫瘤也能偵測到，成為目前肺癌診斷的主要方式。 由於電腦斷層掃描影像資料量大且分析需要大量時間，肺組織中有許多相似腫瘤的結構造成醫生必須逐張檢查，所以判別腫瘤不是一件容易的事，長時間工作除了造成疲憊外還可能會增加錯誤機率，因此，為了減少醫師判別腫瘤的時間，目前急需一套高精度的電腦輔助系統協助醫師判別以增加工作效率以及判別成功率，因此以下研究中我使用捲積神經網路判別腫瘤與非腫瘤期望可以增加判別成功率。 深度學習是目前人工智慧領域中成長最快的領域之一，早期是模仿生物大腦結構的一門科學，人腦重量達一公斤多，結構相當複雜，為了方便電腦模擬所以分成輸入層、輸出層、隱藏層。由於隱藏層可以非常多層所以稱為深度學習，常見的深度學習架構有MLP、CNN、RNN等。並且在語音辨識、視覺辨識、生物醫學等領域取得相當的成果。 傳統的機器學習必須以人工方式擷取影像特徵送入分類器分類，然而，特徵擷取常常不是一件容易的事情，在一些比較複雜的問題上甚至需要整個領域許多年的人力投入。相對於傳統的機器學習演算法，深度學習可以自動在圖片影像中擷取如線段、角、邊、形狀等基礎的特徵，並且找出或組合成更複雜的特徵並且適當分配權重。 LIDC是一個美國國家癌症研究所發起收集的1018個電腦斷層掃描影像數據庫，內容包含肺癌診斷結果以及結節位置標記數據，是由七個學術中心和八家醫學影像公司合作創建，適合開發肺部電腦輔助診斷系統。本論文用LIDC收集的樣本為訓練資料，以卷積神經網路分析其腫瘤的共同特徵，做出這一套分類器，期望可以協助醫師更準確的判斷腫瘤。 本研究使用其中前五百組中厚度為2.5mm的電腦斷層掃描影像樣本，其中腫瘤直徑大於等於3mm的占比約三成五，小於3mm的占比約六成五，結節擷取部分參考LIDC的XML標記資訊，搭配ITK以及openCV擷取影像，接著將資料隨機選為不重複的訓練資料以及測試資料。非結節部分則是從電腦斷層掃描影像中非標記區域隨機擷取。 當結節與非結節部分擷取完後將這些樣本送入卷積神經網路模型中以反向傳播演算法訓練。之後測試由實驗室提供厚度為1.25mm的樣本數據做為比對，此厚度1.25mm的樣本非結節取樣方式與樣本取樣不同，是實驗室以閥值取出較亮區域，再以侵蝕演算法找出，由於厚度不同以及非節節部分取樣方式不同，造成正確率下降，因此再以遷移學習的方式提升正確率。

Tumor is a group of abnormal cells that are detached from normal growth regulation and can then grow without restriction and stay in any process of the cell cycle. Those cells often do not die due to apoptosis and loss of telomerase mechanism. Tumor has become one of the leading causes of death in middle- and high-income countries and in high-income countries. In recent years, cancer has always been the first cause of death published by the Ministry of Health and Welfare. Taking the data of Taiwan’s Ministry of Health and Welfare in the year of 105 as an example, the number of deaths from malignant tumors in Taiwan is 47,760, accounting for 27.7% of the total deaths. However, it is stated in The World Cancer Report published in February 2014 that global cancer mortality will triple by 2035. Subdividing the global mortality rate of malignant tumors, the tracheal, bronchial and lung cancers have always been among the top three categories. However, the five-year survival rate of the third and fourth stages of lung cancer is quite low. Therefore, early diagnosis of lung cancer is the best way to prolong the survival rate of lung cancer. With the development of science and technology, the choice of medical imaging equipment at present is also very diverse, such as chest X-ray, low-dose computed tomography, and positron angiography and so on. Chest X-ray detection application is the earliest, most widely used because of the low cost of equipment and consumables. However, chest X-ray cannot detect tumors less than a centimeter in diameter. Even if the tumor is larger than one cm in diameter, it is easily covered by the trachea, bronchi, thorax, ribs, etc., resulting in blind spots for detection. It is often too late when chest X-ray can detect lung cancer. Computed tomography imaging is a three-dimensional grayscale imaging created by computer reconstruction after X-ray exposure to human body for different tissues have different X-ray absorption rates and different signals can be received by sensors. Compared with chest X-ray imaging, three-dimensional computed tomography imaging has higher resolution and can detect the tumors with a diameter of only 1mm to 2mm, becoming the main method for the diagnosis of lung cancer. However, due to the large amount of computed tomography imaging data and the large amount of time required for analysis, and many similar tumor structures in the lung tissue that make doctors check one by one, it is not an easy task to distinguish tumors. In addition to causing fatigue, long hours of work may increase the chance of error. Therefore, in order to reduce the time for doctors to distinguish tumors, a high-precision computer-aided system is urgently needed to assist doctors in discriminating to increase work efficiency and discriminant success rate. Deep learning is one of the fastest growing fields in the field of artificial intelligence. In the early stage, it is a science that imitates the structure of biological brain. The human brain weighs more than one kilogram, and the structure is quite complex. In order to make computer simulation, it is divided into input layer, output layer and hidden layer. Since the hidden layer can be very multi-layered, it is called deep learning. Common deep learning architectures include MLP, CNN, RNN, and so on. And in the fields of speech recognition, visual identification, biomedical and other fields considerable achievements have been made. In traditional machine learning, imaging features must be extracted manually into the classifier for classification. However, feature extraction is often not an easy task. On some more complicated issues, even the human resources of the entire field for many years are needed. Compared to traditional machine learning algorithms, deep learning can automatically extract some basic features such as line segments, angles, edges, shapes, etc. in picture imaging to find or combine them into more complex features and assign weights appropriately. In this study, the computed tomography imaging samples with a thickness of 2.5 mm in the first 500 sheets will be used. The proportion of tumors with a diameter of 3 mm or more is about 35%, and that of less than 3 mm was about 65%. The nodule extraction part refers to XML tag information of LIDC, with ITK and openCV for extracting images. Then, data is randomly selected as non-repeating training data and test data. The non-nodule part is randomly extracted from the unmarked area of the computed tomography imaging. After the nodule and non-nodule parts have been extracted, these samples are sent to the convolutional neural network model and trained by the back-propagation algorithm. The training period is 120 times. The experimental model consists of a five-layer convolutional layer and a three-layer pooling layer, and the dropout inside the model is set to 0.45, and the laboratory computed tomography imaging data is simultaneously tested. The trained sample classifier testing the LIDC data can achieve an average accuracy of over 95%. Then the sample data provided by the laboratory with a thickness of 1.25mm is tested as the comparison. The non-nodular sampling method of the sample with a thickness of 1.25mm is different from the sample sampling. It is the laboratory to take out the brighter area with the threshold and then find out by the erosion algorithm. Due to the different thickness and different sampling methods of non-nodule parts, although the Sensitivity has a small decline, staying at around 94%, the Specificity only remains over 80%.